Written by the founder of BreezeConvert — a creative director working across set design, game 3D, Unreal/VFX and AI content pipelines. More about the author.
I use AI assistants constantly in my work — summarising briefs, turning messy notes into structure, pulling action items out of long threads. Over a lot of trial and error, one thing became obvious: how you give the model the document matters almost as much as what you ask. Paste a raw Word file's contents or a copy-pasted PDF and the model spends effort untangling formatting before it even reads your question. Give it clean, structured text and the answers get noticeably better. This is the workflow I've settled on.
Why Markdown is the right format
Large language models were trained on enormous amounts of Markdown — it's the plain-text formatting used across the web, documentation, and code. That means the model already "understands" Markdown's small, unambiguous set of cues: # for headings, - for lists, pipes for tables, asterisks for emphasis. When your document arrives as Markdown, the model reads structure directly instead of guessing at it.
Compare that to the alternatives. A raw PDF is a fixed-layout format the model has to reconstruct; copy-pasting from Word drags in smart quotes, stray line breaks, and invisible characters; a screenshot can't be read as text at all. Markdown sidesteps all of it. Tables are the clearest example: paste spreadsheet rows and the columns collapse into a jumble, but a Markdown table keeps every value tied to its header, so the model can actually reason about the data.
Cutting the token count (why it matters)
Everything you send an assistant is measured in tokens — roughly, pieces of words. Two things make this worth caring about. First, every model has a context window, a hard limit on how much it can hold at once; a bloated document can overflow it or crowd out room for the model's answer. Second, on paid tiers you're billed by the token, so waste is literal money.
Clean Markdown is leaner than the raw alternatives because you strip out what the model can't use anyway: embedded images, HTML wrappers, repeated boilerplate, and decorative characters. A useful habit is to watch the token count as you clean — you can often shave 20–30% off just by removing images and normalising junk characters, which is real headroom on a long document. And a note for anyone working in Korean, Chinese, or Japanese: CJK text uses far more tokens per character than English, so those documents hit context limits sooner and benefit most from trimming.
When (and how) to chunk a long document
Sometimes a document is simply bigger than the context window, or you're building a knowledge base for retrieval. That's when you split it into chunks. The wrong way is to cut every N characters — you end up slicing sentences and tables in half, and the model gets confused fragments. The right way is to split on heading boundaries, so each chunk is a complete, self-contained section. Then feed the chunks one at a time, or index them for a retrieval (RAG) system.
A practical target is 4,000–8,000 tokens per chunk for general use, with each piece kept under whatever your assistant's context allows. If a single section is still too big, break it at paragraph boundaries as a fallback. The point is that the model should always receive coherent, whole ideas rather than arbitrary slices.
Give the model an instruction, not just a dump
Once the text is clean, the prompt around it does a lot of work. Vague requests get vague answers. Be specific about the job: "summarise the key risks in this contract," "extract every action item as a checklist," "turn these slides into three talking points," "explain this section in plain language." Putting the instruction before the document and being explicit about the output format you want (a list, a table, a short paragraph) reliably produces more useful results than pasting the text and typing "summarise this."
The part people forget: don't upload confidential files
Here's the tension that started me down this path. The documents most worth handing to an AI — contracts, client briefs, internal strategy, unreleased work — are exactly the ones you shouldn't upload to a random web converter to turn into Markdown first. Many "document to Markdown" tools send your file to their server to process it, which quietly defeats the point of being careful.
The fix is to convert locally. Our Document to Markdown tool runs entirely in your browser — Word, Excel, PowerPoint, 한글, PDF and more become clean Markdown on your own machine, with a token count, one-click prompt buttons, and a chunk splitter built in, and nothing is uploaded. You paste only the finished text into the assistant, on your own terms. For anything sensitive, that's the whole difference.
The quick checklist
Convert the document to clean Markdown; turn on the "plain text" option to strip images and junk; glance at the token count and trim if it's near your limit; chunk on headings if it's genuinely huge; put a specific instruction before the text; and do the conversion locally so the file never leaves your device. Do that, and you'll get better answers for less — and keep the private stuff private.